ERL, COM, and RL with DOE

ERL, COM, and RL with DOE

Richard Mellitz, Samtec

8/30/2017

IEEE 802.3 50 Gb/s, 100 Gb/s, and 200 Gb/s Ethernet Task Force Ad Hoc

ToC

2IEEE 802.3 50 Gb/s, 100 Gb/s, and 200 Gb/s Ethernet Task Force Ad Hoc

Design of experiment (DOE) for simulation

ERL prediction from package parameters

DOE for channels with package parameters

Analysis of Channel DOE data

Fit example to understand some basic terms in MATLAB

R-squared is used determent the quality ofcorrelation• It is called the “coefficient of determination”• R2 = 1 is the best • It is the proportion of the x variable variance

which is predictable.• In this example R2 is 0.9206• i.e. data is 92% correlated to the fit

RMSE is the RMS error• one way to interpret is:

• The equation is, on the average, +/- 0.9804 accurate

• The fit is predicting signal values between 8 and -6.

The fit is an equation called f(x) at left• With fitted p1, p2, and p3 coefficients when are

nearly the same as in the MATLAB code above

IEEE 802.3 50 Gb/s, 100 Gb/s, and 200 Gb/s Ethernet Task Force Ad Hoc 3

MATLAB Code Examplef = @(x) 0.05*(x).^2-.5*(x) -2.2 ;x=-10:.05:10;noise = randn(1,length(x)*1) - .5;signal = f(x)+noise;cftool(x,signal)

X is the independent parameters variablesY is the dependent parameter or in our case simulated COM results

Better fit does not have zero uncertainty

Noise in the experiment isreduced by a factor of 10• In the case 99.9 % correlated to

the fit

RMSE is not zero!• one way to interpret is:

• The equation, is on the average, +/-0.1 accurate

• The fit is predicting signal values between 8 and -6.


MATLAB Code Examplef = @(x) 0.05*(x).^2-.5*(x) -2.2 ;x=-10:.05:10;noise = randn(1,length(x)*1) - .5;signal = f(x)+noise*0.1;cftool(x,signal)X is the independent parameters variables

Y is the dependent parameter or in our case simulated COM results

View the same data in JMP


Fit equation (red) plotted with data

This equation is complicated for more than 1 “x” variable

Fit equation plotted vs data is graphic representation for the ability to predict results base on new x data

A perfect plot would be a 45 degree line

Fit quality data as shown in previous slides

Plot of predicted y verses the new “x” values

Normally there are number of x valuables

The graph is useful for interactively or by inspection determining the effect of moving an “x” variable one way or another

Predicting simulation for millions of combinations of “x” values

From the last slide we have a prediction formula for y as function of x.

If we assign a distribution of for “x” we can determine the resulting y distribution

In this case we are doing a 1 million cases


1 million cases

x distribution

Resulting y distribution

We use linked distributions to determine the effect of the range of x on y

In this case the all the value of x are between -1 and 1 are selected

The dark areas in y correspond the selected cases of x


Multivariate plots are a rough first step

The diagonal is the variable name• In this case x and y

The off diagonal is x plotted vs y and vice versa.• For more variables all possible plots

(linearly) taken 2 at a time are plotted

The red ellipses are the linear correlation factors

Multivariate plots can suggest what to investigate


Using DOE methods to perform simulation prediction

COM, RL, & ERL consensus meeting 9

Determine x parameters

Simulate y parametersfor the collection of x

parameters

Determine collection of values for x parameter

Perform a fit for the collection of x

parameters and simulated y parameters(determine fit equation)

Use fit equation to predict performance for a

much larger range of x parameter values

Checking expectations

COM, RL, & ERL consensus meeting 10

Determine x parameters

Simulate y parametersfor the collection of x

parameters

Determine collection of values for x parameter

Perform a fit for the collection of x

parameters and simulated y parameters(determine fit equation)

Use fit equation to predict performance for a

much larger range of x parameter values

Do the results make

sense?

Effective Return Loss Experiment

X variables are COM package parameters centered on D2.1 COM table

Y is the computed ERL for the specified Zt

From mellitz_060717_3cd_02_adhoc

x yZc

OhmsRd

OhmsZp

OhmsZt

OhmsCd

1e-10 FCp

1e-10 FERLdB

95 50 30 50 1.8 1.1 -9.095 50 30 50 1.8 1.1 -9.096 51 30 50 1.8 1.1 -8.996 51 30 50 1.8 1.1 -8.994 48 30 50 1.8 1.1 -9.194 48 30 50 1.8 1.1 -9.185 45 12 55 2 0.9 -7.995 55 30 55 1.6 1.3 -8.295 45 30 50 1.6 0.9 -10.585 45 30 55 2 1.3 -8.0

105 45 12 45 1.6 1.3 -8.6105 55 12 45 2 1.3 -6.9

85 50 30 55 1.8 0.9 -9.7105 55 30 45 2 0.9 -8.8105 45 30 45 1.8 1.1 -9.4

85 45 12 45 1.6 0.9 -9.985 55 12 55 2 1.3 -6.485 55 12 45 2 0.9 -8.185 45 30 45 2 0.9 -10.2

105 50 12 55 2 1.1 -7.0105 45 12 45 2 0.9 -8.8

95 50 30 45 2 1.3 -8.285 45 30 45 1.6 1.3 -9.6

105 55 30 50 1.8 1.3 -7.8105 45 30 55 1.6 1.3 -8.3105 55 12 55 1.6 0.9 -8.3

85 50 12 50 1.6 1.3 -7.995 55 21 55 2 0.9 -8.185 45 21 55 1.6 1.1 -8.9

105 45 30 55 2 0.9 -8.9105 45 21 50 2 1.3 -7.6

95 55 12 45 1.6 1.1 -8.385 55 12 55 1.6 0.9 -8.785 55 30 50 2 1.1 -8.5

105 50 21 45 1.6 0.9 -9.885 55 21 45 1.8 1.3 -8.095 45 12 55 1.8 1.3 -7.285 45 12 45 2 1.3 -7.785 55 30 45 1.6 0.9 -10.5

105 45 12 55 1.6 0.9 -8.7IEEE 802.3 50 Gb/s, 100 Gb/s, and 200 Gb/s Ethernet Task Force Ad Hoc 11

http://www.ieee802.org/3/cd/public/adhoc/archive/mellitz_060717_3cd_02_adhoc.pdf

ERL fit is very closely tied to package parameters . RMS err is 0.026 dB

(-22.9371813122759) + 0.0180921056094042 * Zc + 0.0568325927712985 * Rd +-0.0589245893510021 * Zp + 0.0680408952372635 * Zt + 2.29551955640816 * Cd+3.2956494591755 * Cp + (Zc - 94.5) * ((Zc - 94.5) * -0.000232738951785374) + (Zc - 94.5) * ((Rd - 49.575) * 0.000677332168959554) + (Zc - 94.5) * ((Zp-22.125) * 0.00111325213652863) + (Zc - 94.5) * ((Zt - 49.875) *0.000238892775679064) + (Rd - 49.575) * ((Zt - 49.875) * -0.00396965785082018) + (Zp - 22.125) * ((Zt - 49.875) * -0.00103701659658663) + (Zt - 49.875) * ((Zt-49.875) * 0.00196432071730769) + (Zc - 94.5) * ((Cd - 1.805) *-0.0153551291854686) + (Rd - 49.575) * ((Cd - 1.805) * 0.0130086172404367) + (Zp - 22.125) * ((Cd - 1.805) * -0.00883864973187424) + (Zt - 49.875) * ((Cd - 1.805) * -0.00345591899210246) + (Zc - 94.5) * ((Cp - 1.095) * -0.0130281234843095) + (Rd - 49.575) * ((Cp - 1.095) * -0.00811355551678069) + (Zp - 22.125) * ((Cp -1.095) * 0.0321656771693624) + (Cd - 1.805) * ((Cp - 1.095) * -0.841281482352236) + (Cp - 1.095) * ((Cp - 1.095) * -1.47094118647978)IEEE 802.3 50 Gb/s, 100 Gb/s, and 200 Gb/s Ethernet Task Force Ad Hoc 12

ERL Prediction Equation

What is ERL for the COM Package


For 30 mm package (Zp) ERL is 9 dB ( ERL dB is negated here)

For 12 mm package (Zp) ERL is 7.9 dB

A simple RL spec could be ERL > 7.9 dB • But is this good enough alone?

• Or is it even helpful?

So all we can say so far is

14COM, RL, & ERL consensus meeting

ERL can be accurate predicted for COM package parameters

Good for determine incremental device/package design improvements because ERL is number not a vector or numbers

ERL is tightly correlated to physical interconnect design choices

Next let's look at COM and ERL for package and channels• Note: Av is adjusted for transmitter parameters

• Av and Afe = 0.004.*Rd+.215;

• Ane = 0.006.*Rd+.304;

Channel Key for KR


number Channel designator

1 '5F3N--Ch1_10_5F3N_t

2 'TEC_STRADAWhisper11p75in_Meg6_Channel_IEEE802_3_cd_Cu_07282016--TEC_Whisper11p75in_THRU_G14G15-07212016

3 'mellitz_01_021716_10dB_6_channels--PAM4_2conn_MP_v2_100ohm_10dB_Nom_thru

4 'mellitz_01_021716_10dB_6_channels--PAM4_2conn_MP_v2_100ohm_10dB_HzLzHz_thru

5 'mellitz_01_021716_10dB_6_channels--PAM4_2conn_MP_v2_100ohm_10dB_LzHzLz_thru




9 '5F3N--Ch4_20_5F3N_t

10 'TEC_STRADAWhisper27in_Meg6_Channel_IEEE802_3_cd_Cu_07282016--TEC_Whisper27in_THRU_G14G15_07202016







17 '5F3N--Ch8_30_5F3N_t

18 'TEC_STRADAWhisper40in_Meg6_Channel_IEEE802_3_cd_Cu_07282016--TEC_Whisper40in_THRU_G14G15_07202016







25 '20dB_HghZ--20dB_HighZ_thru

26 '20dB_HghZ_Nom_HighZ--20dB_HighZ_Nom_HighZ_thru

27 '30dB_HighZ--30dB_HighZ_thru

“X” table for each of the 27 channels


Zc Tx Zc Rx Rd Tx Rd Rx Cp Cd Zt Zp

105 105 45 55 1.1 1.8 50 30

95 95 45 55 1.1 1.8 50 30

95 105 50 50 1.1 1.8 50 30

105 85 55 50 1.1 1.8 50 30

105 85 45 55 1.1 1.8 50 30

85 105 45 55 1.1 1.8 50 30

105 105 55 45 1.1 1.8 50 30

85 85 55 45 1.1 1.8 50 30

85 85 45 50 1.1 1.8 50 30

105 95 45 45 1.1 1.8 50 30

105 85 50 45 1.1 1.8 50 30

85 105 55 45 1.1 1.8 50 30

85 85 50 55 1.1 1.8 50 30

95 85 45 45 1.1 1.8 50 30

95 85 55 55 1.1 1.8 50 30

105 95 51.05 55 1.1 1.8 50 30

105 105 45 45 1.1 1.8 50 30

85 105 55 55 1.1 1.8 50 30

85 105 45 45 1.1 1.8 50 30

105 105 55 55 1.1 1.8 50 30

85 95 55 50 1.1 1.8 50 30

95 95 50 50 1.1 1.8 50 30

94 94 49 49 1.1 1.8 50 30

96 96 51 51 1.1 1.8 50 30

94 94 51 51 1.1 1.8 50 30

96 96 49 49 1.1 1.8 50 30

Highlight COM between 3dB and 4 dBand channel ERL > 8 dB


Failing channel are failing if COM decreases

Channels passing with great margin are not interesting here.

Channel COM ERL11 ERL22

com min from D2.1 delta

'5F3N--Ch1_10_5F3N_t 6.07 -10.69 -11.62 0.49

'TEC_STRADAWhisper11p75in_Meg6_Channel_IEEE802_3_cd_Cu_07282016--TEC_Whisper11p75in_THRU_G14G15-07212016 6.75 -13.76 -13.34 0.23

'mellitz_01_021716_10dB_6_channels--PAM4_2conn_MP_v2_100ohm_10dB_Nom_thru 5.25 -8.79 -5.68 0.47

'mellitz_01_021716_10dB_6_channels--PAM4_2conn_MP_v2_100ohm_10dB_HzLzHz_thru 5.53 -8.98 -5.36 0.56

'mellitz_01_021716_10dB_6_channels--PAM4_2conn_MP_v2_100ohm_10dB_LzHzLz_thru 4.57 -7.11 -4.94 0.61




'5F3N--Ch4_20_5F3N_t 5.60 -10.31 -13.27 0.34

'TEC_STRADAWhisper27in_Meg6_Channel_IEEE802_3_cd_Cu_07282016--TEC_Whisper27in_THRU_G14G15_07202016 4.78 -14.48 -13.71 0.15







'5F3N--Ch8_30_5F3N_t 3.07 -11.25 -13.76 0.28

'TEC_STRADAWhisper40in_Meg6_Channel_IEEE802_3_cd_Cu_07282016--TEC_Whisper40in_THRU_G14G15_07202016 1.68 -14.90 -14.08 0.18







'20dB_HghZ--20dB_HighZ_thru 3.15 -17.17 -16.71 0.37

'20dB_HghZ_Nom_HighZ--20dB_HighZ_Nom_HighZ_thru 3.27 -18.95 -18.45 0.40

'30dB_HighZ--30dB_HighZ_thru 3.16 -17.34 -17.08 0.25

Return loss for selected channel (3 dB to 4 dB COM)


2 channels correspond to limit crossing pass com

Results have interesting results

ERL and COM in dB are the individual y axes.

Zc and Rx (ohms) are the individual x axes.

Does this make sense?

No

Further investigation found a problem with bifurcation of Tx and Rx Rd in the COM MatLab

Does not affect any computation made withCOM table used in standards so far

IEEE 802.3 50 Gb/s, 100 Gb/s, and 200 Gb/s Ethernet Task Force Ad Hoc19

COM 2.0b results rectifies problem


Variability still exists irrespective of return loss or ERL

1 Million combination of Rd and Zc

Limiting ER does not affect downward COM distribution limit very much


Conclusion


Return loss and/or ERL should not be used for channels

Return loss and/or ERL should not be used for devices

Suspected reason: • Return loss can have a constructive or destructive impact

Something like VSWR (voltage standing wave ratio) might be a better discriminator.

Documents

ERL, COM, and RL with DOE